Manufacturing Downtime & Maintenance Cost Analysis
Which machines fail, what does the downtime cost, and where should the maintenance budget actually go? Analysis of 10,000 machining cycles: failure-mode KPIs, a documented downtime cost model, an Excel report for stakeholders, and a Power BI dashboard spec. This is the analysis every maintenance meeting wants and rarely has.
Why this project
I trained as a mechanical engineer and spent nine years around industrial equipment before moving into data. Maintenance budgets are usually split evenly across failure categories because nobody has quantified where the money actually goes. This project does that quantification on the AI4I 2020 dataset, and it is deliberately the descriptive sibling of my predictive maintenance MLOps project: same dataset, same engineering context. Analysis first, prediction second, one narrative.
What the data says
- 3.39% of cycles end in failure, costing an estimated £2.33M in lost margin and parts in the window analysed.
- Two failure modes carry 70% of the cost. Overstrain (39%) and power failures (30%) pair high frequency with the longest repairs; tool wear is frequent but cheap (5%).
- The low-spec (L) line is the reliability problem: it fails at 39.2 per 1,000 cycles, nearly twice the H line's rate, and carries the highest cost per 1,000 cycles.
- Four of the five failure modes have clear sensor signatures, so they are preventable: every heat-dissipation failure happened with a cooling gap under 8.6 K, every power failure sat outside the 3.5-9 kW band, and a tool change-out at ~204 minutes prevents roughly 90% of tool-wear failures.
- A flat budget split is the wrong split: cost-share allocation moves ~19 points of budget into overstrain prevention and ~15 points out of tool-wear response.
How it works
Measured vs modelled, kept separate
Failure counts and sensor signatures are measured from the data. Pound figures are estimates from a documented cost model, labelled as estimates everywhere they appear. The mode ranking is robust to the assumptions; the absolute figures are not; and the analysis never mixes the two. That distinction is the difference between analysis a finance director can trust and a chart that gets picked apart in the first five minutes.
Skills demonstrated
- Cost analysis and KPI design (failure rate, MTTR, downtime, cost per 1,000 cycles)
- Python analysis with a tested, reproducible pipeline (runs fully in CI)
- Excel deliverables: formatted multi-sheet KPI workbook generated with openpyxl
- Power BI: star schema, DAX measures including a Pareto, dashboard spec
- Mechanical engineering domain knowledge applied to data work